On Thu, Sep 02, 2021 at 13:51:26 +0100, Peter Maydell wrote:
> On Thu, 2 Sept 2021 at 13:43, Leif Lindholm <l...@nuviainc.com> wrote:
> > On Thu, Aug 19, 2021 at 14:27:19 +0100, Peter Maydell wrote:
> > > If you want a command line switch to let the user say whether the
> > > ITS should be present or not, you should use the same thing the
> > > virt board does, which is a bool property "its".
> > >
> > > If you want the sbsa-ref board to become a proper "versioned machine
> > > type" such that users can say "-M sbsa-ref-6.1" and get the SBSA
> > > board exactly as QEMU 6.1 supplied it, that looks completely different
> > > (and is a heavy back-compatibility burden, so needs discussion about
> > > whether now is the right time to do it), and probably is better not
> > > tangled up with this patchseries.
> >
> > Hmm. I mean, you're absolutely right about the complexity and need for
> > discussion. My concern is that we cannot insert the ITS block in the
> > memory map where it would be in an ARM GIC implementation without
> > changing the memory map (pushing the redistributor further down).
> >
> > It is also true that simply versioning sbsa-ref does not mean we end
> > up with a properly aligned with ARM IP register frame layout. Some
> > additional logic is required for that.
> >
> > If we assume that we don't want to further complicate this set by
> > adding the additional logic *now*, I see three options:
> > - Implement memory map versioning for sbsa-ref for this set, placing
> > the ITS (if enabled) directly after the DIST for sbsa-ref-6.2.
> > - In this set, place the ITS frames in a different location relative
> > to the REDIST frames than it will end up once the further logic is
> > implemented.
> > - Drop the sbsa-ref ITS support from this set, and bring it in with
> > the set implementing the additional logic.
>
> I don't think implementing versioned machine types helps you much
> anyway, because your users are all going to be currently using
> -M sbsa-ref, so they will by default see the change in device layout.
>
> I do think that we should get the "with an ITS" device layout right
> from the start, so that we're only dealing with 2 possibilities
> (what we have today, and the final intended layout), not 3 (today,
> final layout, some intermediate thing).
>
> How does guest software on this board figure out the memory
> layout ? Is there a mechanism for telling it "this is version 2
> of this board" ?
Not currently. Aiming to look like most SBSA platforms, it is
hard-wired, and firmware passes that information to the os.
This is the kind of thing I eventually want to do using a PV-model
SCP. As a stop-gap, we could push it through the DT, like we do for
cpus and memory?
/
Leif
